Professional Experience

Key Achievements

• Architected production-grade MLOps pipeline processing 180 satellite images into 12,000+ training tiles with overlap-aware tiling and zero data leakage across train/val/test splits
• Trained three complementary models: YOLO11m object detection (mAP50 0.836), YOLO11m instance segmentation (mAP50-mask 0.737), and fine-tuned SAM1 ViT-H semantic segmentation (mIoU 0.804) covering 10 object classes
• Fine-tuned SAM1 ViT-H mask decoder with pre-computed embedding caching across 9,958 tiles, reducing epoch time from 88 min to 5 min on T4 GPU using custom Dice + Focal loss and mixed-precision training
• Exported inference results as GeoJSON FeatureCollections with WGS84 coordinates, enabling visualization in Google Earth, QGIS, Leaflet, and any GIS-compatible tool